Hindawi Publishing Corporation Abstract and Applied Analysis Volume 2014, Article ID 430932, 14 pages http://dx.doi.org/10.1155/2014/430932
Research Article The Flattened Aggregate Constraint Homotopy Method for Nonlinear Programming Problems with Many Nonlinear Constraints Zhengyong Zhou1 and Bo Yu2 1 2
School of Mathematics and Computer Sciences, Shanxi Normal University, Linfen, Shanxi 041004, China School of Mathematical Sciences, Dalian University of Technology, Dalian, Liaoning 116024, China
Correspondence should be addressed to Zhengyong Zhou;
[email protected] Received 8 February 2014; Accepted 27 April 2014; Published 29 May 2014 Academic Editor: Victor Kovtunenko Copyright Β© 2014 Z. Zhou and B. Yu. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The aggregate constraint homotopy method uses a single smoothing constraint instead of π-constraints to reduce the dimension of its homotopy map, and hence it is expected to be more efficient than the combined homotopy interior point method when the number of constraints is very large. However, the gradient and Hessian of the aggregate constraint function are complicated combinations of gradients and Hessians of all constraint functions, and hence they are expensive to calculate when the number of constraint functions is very large. In order to improve the performance of the aggregate constraint homotopy method for solving nonlinear programming problems, with few variables and many nonlinear constraints, a flattened aggregate constraint homotopy method, that can save much computation of gradients and Hessians of constraint functions, is presented. Under some similar conditions for other homotopy methods, existence and convergence of a smooth homotopy path are proven. A numerical procedure is given to implement the proposed homotopy method, preliminary computational results show its performance, and it is also competitive with the state-of-the-art solver KNITRO for solving large-scale nonlinear optimization.
1. Introduction In this paper, we consider the following nonlinear programming problem: min
π (π₯) ,
s.t. π (π₯) β€ 0,
(1)
where π₯ β π
π is the variable, π(π₯) = (π1 (π₯), . . . , ππ (π₯))π , π : π
π β π
, and ππ (π₯) : π
π β π
, π = 1, . . . , π, are three times continuously differentiable, and π is very large, but π is moderate. It has wide applications, and a typical situation is the discretized semi-infinite programming problem. From the mid-1980s, much attention has been paid to interior point methods for mathematical programming, and many results on theory, algorithms, and applications on linear programming, convex programming, complementarity problems, semidefinite programming, and linear cone programming were obtained (see monographs [1β6] and
references therein). For nonlinear programming, the typical algorithms used were the Newton-type methods to the perturbed first-order necessary conditions combined with line search or trust region methods with a proper merit function (e.g., [7β9]). The general conditions of global convergence for these methods required that the feasible set be bounded and that the Jacobian matrix be uniformly nonsingular. Another typical class of globally convergent methods for nonlinear programming was probability-one homotopy methods (e.g., [10β12]), whose global convergence can be established under some weaker conditions than the ones for Newton-type methods. The excellent feature is that, unlike line search or trust region methods, they do not depend on the descent of a merit function and so are insensitive to the local minimum of the merit function, in which any search direction is not a descent direction of the merit function. In [10, 11], Feng et al. proposed a homotopy method for nonlinear programming (1), which was called the combined homotopy interior point (abbreviated by CHIP) method;
2
Abstract and Applied Analysis
its global convergence was proven under the normal cone condition (see below for its definition) for the feasible set as well as some common conditions. On the basis of the CHIP method, some modified CHIP methods were presented in [13, 14]; the global convergence was established under the quasinormal cone and pseudocone condition for the feasible set, respectively. In [12], Watson described some probabilityone homotopy methods for the unconstrained and inequality constrained optimization, whose global convergence was established under some weaker assumptions. Recently, Yu and Shang proposed a constraint shifting combined homotopy method in [15, 16], in which not only the objective function but also the constraint functions were regularly deformed. The global convergence was proven under the condition that the initial feasible set, which approaches the feasible set of (1) as the homotopy parameter changes from 1 to 0, not necessarily the feasible set of (1), satisfies the normal cone condition. Let πmax (π₯) = max1β€πβ€π {ππ (π₯)}; then (1) is equivalent to min s.t.
π (π₯) , πmax (π₯) β€ 0,
(2)
which has only one, but nonsmooth, constraint. In [17], the following aggregate function was introduced, which is a smooth approximation of πmax (π₯) with a smoothing parameter π‘ > 0 and induced from the max-entropy theory: π
πΜ (π₯, π‘) = π‘ ln β exp ( π=1
ππ (π₯) ). π‘
(3)
It is also known as exponential penalty function (see [18]). By using it for all constraint functions of problem (1), an aggregate constraint homotopy (abbreviated by ACH) method was presented by Yu et al. in [19], whose global convergence was obtained under the condition that the feasible set satisfies the weak normal cone condition. Although the ACH method has only one smoothing constraint, the gradient and Hessian of the aggregate constraint function π
βπ₯ πΜ (π₯, π‘) = βππ (π₯, π‘) βππ (π₯) , π=1
π ππΜ (π₯, π‘) 1 = (πΜ (π₯, π‘) β βππ (π₯, π‘) ππ (π₯)) , ππ‘ π‘ π=1 2 πΜ (π₯, π‘) βπ₯π₯ π
= βππ (π₯, π‘) β2 ππ (π₯) π=1
1π π + βππ (π₯, π‘) βππ (π₯) (βππ (π₯)) π‘ π=1 π 1π π β βππ (π₯, π‘) βππ (π₯) βππ (π₯, π‘) (βππ (π₯)) , π‘ π=1 π=1
2 πΜ (π₯, π‘) βπ₯π‘
=
π 1 π ( π ππ (π₯, π‘) ππ (π₯) π‘) βπ β β (π₯, (π₯) π π‘2 π=1 π π=1 π
ββππ (π₯, π‘) ππ (π₯) βππ (π₯)) , π=1
(4) where ππ (π₯, π‘) =
exp (ππ (π₯) /π‘) β (0, 1) , exp (ππ (π₯) /π‘)
βπ π=1
π
βππ (π₯, π‘) = 1, π=1
(5) are complicated combinations of gradients and Hessians of all constraint functions and, hence, are expensive to calculate when π is very large. Throughout this paper, we assume that the nonlinear programming problem (1) possesses a very large number of nonlinear constraints, but a small number of variables, and the objective and constraint functions are not sparse. For such a problem, the number of constraint functions can be so large that the computation of gradients and Hessians of all constraint functions is very expensive and cannot be stored in memory and, hence, the general numerical methods for solving nonlinear programming are not efficient. Although active set methods only need to calculate gradients and Hessians of a part of constraint functions, require lower storage, and have faster numerical solution, the working set is difficult to estimate without knowing the internal structure of the problem. In this paper, we present a new homotopy method called the flattened aggregate constraint homotopy (abbreviated by FACH) method for nonlinear programming (1) by using a new smoothing technique, in which only a part of constraint functions is aggregated. Under the normal cone condition for the feasible set and some other general assumptions, we prove that the FACH method can determine a smooth homotopy path from a given interior point of the feasible set to a KKT point of (1), and preliminary numerical results demonstrate its efficiency. The rest of this paper is organized as follows. We conclude this section with some notations, definitions, and a lemma. The flattened aggregate constraint function with some properties is given in Section 2. The homotopy map, existence and convergence of a smooth homotopy path with proof are given in Section 3. A numerical procedure for tracking the smooth homotopy path, and numerical test results with some remarks are given in Section 4. Finally, we conclude the paper with some remarks in Section 5. When discussing scalars and scalar-valued functions, subscripts refer to iteration step so that superscripts can be used for exponentiation. In contrast, for vectors and vectorvalued functions, subscripts are used to indicate components, whereas superscripts are used to indicate the iteration step. The identity matrix is represented by πΌ. Unless otherwise specified, β β
β denotes the Euclidean norm. Ξ© = {π₯ β π
π | π(π₯) β€ 0} is the feasible set of (1),
Abstract and Applied Analysis
3
whereas Ξ©0 = {π₯ β π
π | π(π₯) < 0} is the interior of Ξ© and πΞ© = Ξ© \ Ξ©0 is the boundary of Ξ©. The symbols π
+π π and π
++ denote the nonnegative and positive quadrants of π π
, respectively. The active index set is denoted by πΌ(π₯) = {π β {1, . . . , π} | ππ (π₯) = 0} at π₯ β Ξ©. For a function πΉ : π
π β π
π , βπΉ(π₯) is the π Γ π matrix whose (π, π)th element is ππΉπ (π₯)/ππ₯π ; πΉβ1 (π) = {π₯ | πΉ(π₯) β π} is the setvalued inverse for the set π β π
π . Definition 1 (see [10]). The set Ξ© satisfies the normal cone condition if and only if the normal cone of Ξ© at any π₯ β πΞ© does not meet Ξ©0 ; that is, 0
{π₯ + β π¦π βππ (π₯) | π¦π β₯ 0} β© Ξ© = 0.
(6)
constraints, which is weaker than the linear independence constraint qualification. The last assumption, the normal cone condition of the feasible set, is a generalization of the convex condition for the feasible set. Indeed, if Ξ© is a convex set, then it satisfies the normal cone condition. Some simple nonconvex sets, satisfying the normal cone condition, are shown in [10]. In this paper, we construct a flattened aggregate constraint function as follows: πΜ(π,π1 ,π2 ,πΌ) (π₯, π‘) π
= ππ‘ ln (βπ (ππ (π₯) , π‘) exp ( π=1
πβπΌ(π₯)
Definition 2. If and only if there exists πβ β π
+π such that (π₯β , πβ ) satisfies βπ (π₯β ) + βπ (π₯β ) πβ = 0, πβπ ππ
β
πβπ
(π₯ ) = 0,
β
β₯ 0, ππ (π₯ ) β€ 0, π = 1, . . . , π,
Definition 3. Let π β π
π be an open set and let πΉ : π β π
π be a differentiable map. We say π¦ β π
π is a regular value of πΉ if and only if Range [
ππΉ (π₯) ] = π
π , ππ₯
βπ₯ β πΉβ1 (π¦) .
where π (π‘) = π1 π‘ + π2 ,
(7)
then π₯β is called a KKT point of (1) and πβ is the corresponding Lagrangian multiplier.
(8)
Lemma 4 (parameterized Sard theorem [20]). Let π β π
π and π β π
π be two open sets and let πΉ : π Γ π β π
π be a πΆπ differentiable map with π > max{0, π β π}. If 0 β π
π is a regular value of πΉ, then, for almost all π β π, 0 is a regular value of πΉπ = πΉ(β
, π).
0, for π§ β€ βπΌπ (π‘) with πΌ > 1, { { π (π§, π‘) = {1, for π§ β₯ βπ (π‘) , { π π‘) , otherwise, (π§, {
π (βπΌπ (π‘) , π‘) = 0,
π (βπ (π‘) , π‘) = 1,
Assumption 5. Ξ©0 is nonempty and Ξ© is bounded.
π2 π (βπ (π‘) , π‘) = 0, ππ§2
β π¦π βππ (π₯) = 0,
πβπΌ(π₯)
π¦π β₯ 0, π β πΌ (π₯) σ³¨β π¦π = 0, π β πΌ (π₯) . (9)
Assumption 7. Ξ© satisfies the normal cone condition (see Definition 1). The first assumption is the Slaterβs condition and the boundedness of the feasible set, which are two basic conditions. The second assumption provides the regularity of
(12)
where, for any π‘ β [0, 1], π(π§, π‘) satisfies
In this section, we suppose that the following assumptions hold.
Assumption 6. For any π₯ β πΞ©, {βππ (π₯) | π β πΌ(π₯)} are positive independent; that is,
(11)
π1 > 0, π2 > 0, 0 < π β€ 1 are some adjusting parameters, π(π§, π‘) : π
Γ[0, 1] β [0, 1] is a πΆ3 bivariate function satisfying π(π§, π‘) = 0 for π§ β€ βπΌπ(π‘) with πΌ > 1, and π(π§, π‘) = 1 for π§ β₯ βπ(π‘). For simplicity of discussion, we write πΜ(π,π1 ,π2 ,πΌ) (π₯, π‘) Μ π‘). By its definition, π(π§, π‘) can be expressed as as π(π₯,
π2 π (βπΌπ (π‘) , π‘) = 0, ππ§2
2. The Flattened Aggregate Constraint Function
ππ (π₯) π (π‘) ) + exp (β )) , ππ‘ ππ‘ (10)
ππ (βπΌπ (π‘) , π‘) = 0, ππ§ π3 π (βπΌπ (π‘) , π‘) = 0, ππ§3 ππ (βπ (π‘) , π‘) = 0, ππ§
(13)
π3 π (βπ (π‘) , π‘) = 0. ππ§3
There exist many kinds of functions π(π§, π‘), such as polynomial functions and spline functions. Since polynomial functions have simple mathematical expressions and many excellent properties, throughout this paper, π(π§, π‘) is chosen as π (π§, π‘) =
β20(π§ + π (π‘))7 β70(π§ + π (π‘))6 + (πΌπ (π‘) β π (π‘))7 (πΌπ (π‘) β π (π‘))6 β84(π§ + π (π‘))5 β35(π§ + π (π‘))4 + + + 1. (πΌπ (π‘) β π (π‘))5 (πΌπ (π‘) β π (π‘))4
(14)
4
Abstract and Applied Analysis
Let πΌπ (π₯, π‘) = {π | ππ (π₯) > βπΌπ(π‘)}; by the definition of Μ π‘) in (10) can be rewritten as π(π§, π‘), π(π₯,
1 β₯ β π π (π₯, π‘)
π (π₯) ) πΜ (π₯, π‘) = ππ‘ ln ( β π (ππ (π₯) , π‘) exp ( π ππ‘ πβπΌ (π₯,π‘) π
+ exp (β
(2) By limπ₯βπ₯,π‘βπ‘ (βπ(π‘) β ππβ (π₯)) = βπ(π‘) < 0, we have
πβπΌπ (π₯,π‘)
(15) π (π‘) )). ππ‘
=
πβπΌπ (π₯,π‘)
(16)
βπβπΌπ (π₯,π‘) π (ππ (π₯) , π‘) exp (ππ (π₯) /ππ‘) + exp (βπ (π‘) /ππ‘)
β₯
π (ππβ (π₯) , π‘) exp (ππβ (π₯) /ππ‘) π (ππβ (π₯) , π‘) exp (ππβ (π₯) /ππ‘) + exp (βπ (π‘) /ππ‘)
β₯
π (ππβ (π₯) , π‘) 1 + exp ((βπ (π‘) β ππβ (π₯)) /ππ‘)
Μ π‘) with respect to π₯ is The gradient of π(π₯, βπ₯ πΜ (π₯, π‘) = β (π π (π₯, π‘) + ππ (π₯, π‘)) βππ (π₯) ,
βπβπΌπ (π₯,π‘) π (ππ (π₯) , π‘) exp (ππ (π₯) /ππ‘)
σ³¨β 1. (20)
where π π (π₯, π‘) =
π (ππ (π₯) , π‘) exp (ππ (π₯) /ππ‘)
, βπβπΌπ (π₯,π‘) π (ππ (π₯) , π‘) exp (ππ (π₯) /ππ‘) + exp (βπ (π‘) /ππ‘) (17)
ππ (π₯, π‘) =
Hence, βπβπΌπ (π₯,π‘) π π (π₯, π‘) β 1. Together with item 1, we have βπβπΌ(π₯) π π (π₯, π‘) β 1 for ππ‘ β 0, π‘ β π‘, and π₯ β π₯. (3) If ππ (π₯) β₯ βπ(π‘), then ππ(ππ (π₯), π‘)/ππ§ = 0 by its definition, and hence ππ (π₯, π‘) = 0;
(21)
else
ππ‘ππ (ππ (π₯) , π‘) /ππ§ exp (ππ (π₯) /ππ‘)
. βπβπΌπ (π₯,π‘) π (ππ (π₯) , π‘) exp (ππ (π₯) /ππ‘) + exp (βπ (π‘) /ππ‘) (18)
σ΅¨ σ΅¨ σ΅¨σ΅¨ σ΅¨ ππ‘ σ΅¨σ΅¨ππ (ππ (π₯) , π‘) /ππ§σ΅¨σ΅¨σ΅¨ exp (ππ (π₯) /ππ‘) β€ ππ‘π, σ΅¨σ΅¨ππ (π₯, π‘)σ΅¨σ΅¨σ΅¨ β€ σ΅¨ exp (βπ (π‘) /ππ‘) (22)
Μ π‘) only relate to a part Then the gradient and Hessian of π(π₯, of constraint functions; that is, π β πΌπ (π₯, π‘). Proposition 8. For any π > 0, π‘ β (0, 1], π₯ β Ξ© with ππ‘ β 0, π‘ β π‘ β [0, 1], π₯ β π₯ β πΞ©,
where π = max(π§,π‘)βπ
Γ[0,1] |ππ(π§, π‘)/ππ§|. Then, we have π(π₯, π‘) β 0 for ππ‘ β 0, π‘ β π‘, and π₯ β π₯. Proposition 9. For any given 0 < π β€ 1,
(1) π π (π₯, π‘) β 0 for π β πΌ(π₯);
Μ π‘) β€ max{πmax (π₯), βπ(π‘)} + ππ‘ ln(ππΌπ + 1), (1) βπ(π‘) β€ π(π₯,
(2) βπβπΌ(π₯) π π (π₯, π‘) β 1;
Μ π‘) β€ πmax (π₯) + ππ‘ ln(ππΌπ + 1) (2) πmax (π₯) β€ π(π₯, when πmax (π₯) β₯ βπ(π‘),
(3) π(π₯, π‘) β 0. Proof. (1) Because π₯ β πΞ©, there exists πβ β πΌ(π₯) such that ππβ (π₯) = 0. By the continuity of π(π§, π‘) and π(π₯), we know that limπ₯βπ₯,π‘βπ‘ π(ππβ (π₯), π‘) = π(ππβ (π₯), π‘) = 1 and limπ₯βπ₯ (ππ (π₯) β ππβ (π₯)) = ππ (π₯) < 0 for π β πΌ(π₯); hence
Μ π‘) = βπ(π‘) when πmax (π₯) β€ βπΌπ(π‘) with πΌ > 1, (3) π(π₯, where ππΌπ denotes the cardinality of πΌπ (π₯, π‘). Proof. (1) By π(π§, π‘) β₯ 0, we have
0 β€ π π (π₯, π‘) exp (ππ (π₯) /ππ‘) β€ π (ππβ (π₯) , π‘) exp (ππβ (π₯) /ππ‘) exp ((ππ (π₯) β ππβ (π₯)) /ππ‘) = π (ππβ (π₯) , π‘) σ³¨β 0. Hence, item 1 is satisfied for ππ‘ β 0, π‘ β π‘, and π₯ β π₯.
πΜ (π₯, π‘) = ππ‘ ln ( β π (ππ (π₯) , π‘) exp ( πβπΌπ (π₯,π‘)
(19)
+ exp (β β₯ ππ‘ ln (exp (β = βπ (π‘) .
π (π‘) )) ππ‘
π (π‘) )) ππ‘
ππ (π₯) ) ππ‘
(23)
Abstract and Applied Analysis
5
Then the left inequality can be obtained. By π(π§, π‘) β€ 1, we have πΜ (π₯, π‘) = ππ‘ ln ( β π (ππ (π₯) , π‘) exp ( πβπΌπ (π₯,π‘)
π β (0, 1] satisfying π β€ β max{πmax (π₯β ), βπ(1)}/(2 ln(π+1)), Μ by Proposition 9(1), for any π₯ β Ξ©, πΜ (π₯, 1) β€ max {πmax (π₯) , βπ (1)} + π ln (π + 1)
ππ (π₯) ) ππ‘
β€ max {πmax (π₯β ) , βπ (1)} + π ln (π + 1)
π (π‘) + exp (β )) ππ‘
=
1 max {πmax (π₯β ) , βπ (1)} 2
(27)
< 0, π (π₯) π (π‘) β€ ππ‘ ln ( β exp ( π ) + exp (β )) . ππ‘ ππ‘ πβπΌ (π₯,π‘) π
(24)
Μ β Ξ©π (1)0 . which means Ξ© (3) For any π₯ β Ξ©, that is, πmax (π₯) β€ 0, by the right inequality of Proposition 9(1), lim πΜ (π₯, π‘) β€ lim max {πmax (π₯) , βπ (π‘)}
Then the right inequality can be obtained by Proposition 2.3 in [19]. (2) For any π₯, π‘ with πmax (π₯) β₯ βπ(π‘), let ππβ (π₯) = πmax (π₯); then π(ππβ (π₯), π‘) = 1 by its definition, and hence
π‘β0
π‘β0
(28)
β€ max {πmax (π₯) , βπ2 } β€ 0, where the second inequality comes from βπ(π‘) = βπ1 π‘ β π2 β€ βπ2 < 0 for π‘ β [0, 1], which means Ξ© β limπ‘β0 Ξ©π (π‘). Then, together with item 1,
π (π₯) πΜ (π₯, π‘) = ππ‘ ln ( β π (ππ (π₯) , π‘) exp ( π ) ππ‘ πβπΌ (π₯,π‘) π
+ exp (β
β₯ ππ‘ ln (π (ππβ (π₯) , π‘) exp ( = ππ‘ ln (exp (
lim Ξ©π (π‘) = Ξ©.
π (π‘) )) ππ‘ ππβ (π₯) )) ππ‘
π‘β0
(25)
ππβ (π₯) )) ππ‘
Proposition 11. There exists a π β (0, 1] such that, for any Μ π‘) =ΜΈ 0. π‘ β (0, 1], π₯ β πΞ©π (π‘); then βπ₯ π(π₯, Proof. If not, for any π β (0, 1], there exist corresponding Μ π , π‘π ) = 0, which π‘π β (0, 1] and π₯π β πΞ©π (π‘π ) such that βπ₯ π(π₯ means that there must exist three sequences {ππ }β π=1 β (0, 1], π β π {π‘π }β β (0, 1], and {π₯ } with π₯ β πΞ© (π‘ ππ π ), such that π=1 π=1 π ππ β 0, π‘π β π‘, π₯ β π₯ β πΞ©, and
= ππβ (π₯) = πmax (π₯) ; then the left inequality is true. The right inequality can be obtained by item 1. Μ π‘). (3) It is trivial by the definition of π(π₯, Μ π‘) β€ 0} and Ξ©π (π‘)0 = Proposition 10. Let Ξ©π (π‘) = {π₯ | π(π₯, Μ π‘) < 0}; one has {π₯ | π(π₯, (1) Ξ©π (π‘) β Ξ© for π‘ β (0, 1]; Μ β Ξ©0 , there exists (2) for any bounded and closed set Ξ© Μ a π β (0, 1], such that Ξ© β Ξ©π (1)0 ; (3) Ξ©π (π‘) β Ξ© as π‘ β 0. Proof. (1) If π₯ β Ξ©, which is equivalent to πmax (π₯) > 0, by Proposition 9(2), 0 < πmax (π₯) β€ πΜ (π₯, π‘) ,
(29)
(26)
which means π₯ β Ξ©π (π‘); then we have Ξ©π (π‘) β Ξ©. Μ is a (2) By the continuity of π(π₯) and the fact that Ξ© β Μ bounded closed set, there exists a point π₯ β Ξ© such that Μ and πmax (π₯β ) < 0. For any πmax (π₯) reaches its maximum in Ξ©,
βπ₯ πΜ (π₯π , π‘π ) =
β (π π (π₯π , π‘π ) + ππ (π₯π , π‘π )) βππ (π₯π ) σ³¨β 0 πβπΌπ (π₯π ,π‘π )
(30)
as π β β. Then, by Proposition 8, we know that π(π₯π , π‘π ) β π, π(π₯π , π‘π ) β 0 as π β β, and β
ππ = β ππ = 1,
πβlimπσ³¨ββ πΌπ (π₯π ,π‘π )
πβπΌ(π₯)
(31)
where the first equality comes from that ππ = 0 for π β πΌ(π₯). Therefore, we have βπβπΌ(π₯) ππ βππ (π₯) = 0 by taking limits on (30); this is a contradiction to Assumption 6. Μ β Ξ©0 , there Proposition 12. For any bounded closed set Ξ© exists a π β (0, 1] such that Μ = 0, {π₯ + πβπ₯ πΜ (π₯, 1) | π > 0} β© Ξ© for any π₯ β πΞ©π (1).
(32)
6
Abstract and Applied Analysis
Μ and four Proof. If not, there exist a bounded closed set Ξ© β π β π sequences {ππ }π=1 β (0, 1], {π₯ }π=1 with π₯ β πΞ©ππ (1), π β π Μ {π₯Μπ }β π=1 β Ξ©, and {π }π=1 > 0 such that ππ β 0, π₯ β π₯ β πΞ©, Μ π π (π₯π , 1) β πβ β₯ 0, ππ (π₯π , 1) β 0 as π β β, π₯Μπ β π₯Μ β Ξ©, π and π₯Μπ = π₯π + ππ
β (π π (π₯π , 1) + ππ (π₯π , 1)) βππ (π₯π ) . πβπΌπ (π₯π ,1) (33)
π» (π₯, π, π‘) (1 β π‘) (βπ (π₯) + πβπ₯ πΜ (π₯, π‘)) + π‘ (π₯ β π₯0 ) ) =( ππΜ (π₯, π‘) β π‘π0 πΜ (π₯0 , 1)
(39)
= 0,
By Proposition 8, we have β
3.1. The Flattened Aggregate Constraint Homotopy. Using Μ π‘) in (10), the flattened aggregate constraint function π(π₯, we construct the following flattened aggregate constraint homotopy:
πβπ = β πβπ = 1,
πβlimπσ³¨ββ πΌπ (π₯π ,1)
πβπΌ(π₯)
(34)
where the first equality comes from that πβπ = 0 for π β πΌ(π₯). Then, we have β (π π (π₯π , 1) + ππ (π₯π , 1)) βππ (π₯π ) πβπΌπ (π₯π ,1)
(35)
σ³¨β β πβπ βππ (π₯) . πβπΌ(π₯)
By Assumption 6 and using (34), β πβπ βππ (π₯) =ΜΈ 0.
πβπΌ(π₯)
(36)
Hence {ππ }β π=1 is a bounded sequence, and there must exist a subsequence converging to π > 0; then we have by (33) π₯Μ = π₯ + π β πβπ βππ (π₯) , πβπΌ(π₯)
(37)
which contradicts Assumption 7.
3. The Flattened Aggregate Constraint Homotopy Method In [19], the following aggregate constraint homotopy was introduced: π» (π₯, π, π‘) (1 β π‘) (βπ (π₯) + πβπ₯ πΜπ (π₯, π‘)) + π‘ (π₯ β π₯0 ) ) (38) =( ππΜπ (π₯, π‘) β π‘π0 πΜπ (π₯0 , 1) = 0, where πΜπ (π₯, π‘) = ππ‘ ln βπ π=1 exp(ππ (π₯)/ππ‘) with π β (0, 1], π is the Lagrangian multiplier of the aggregate constraint Μ Γ π
++ . function πΜπ (π₯, π‘), and (π₯0 , π0 ) is the starting point in Ξ© Under the weak normal cone condition, which is similar to the normal cone condition, it was proved that the ACH determines a smooth interior path from a given interior point to a KKT point. Then, the predictor-corrector procedure can be applied to trace the homotopy path from (π₯0 , π0 , 1) to (π₯β , πβ , 0), in which (π₯β , πβ ) is a KKT point of (1).
where π is the Lagrangian multiplier of the flattened aggregate constraint and (π₯0 , π0 ) is the starting point and can be Μ Γ π
++ . randomly chosen from Ξ© We give the main theorem on the existence and convergence of a smooth path from (π₯0 , π0 , 1) to (π₯β , πβ , 0), in which (π₯β , πβ ) is a solution of the KKT system of (1), and hence the global convergence of our proposal, namely, the FACH method, can be proven. Theorem 13. Suppose that Assumptions 5β7 hold and π β Μβ (0, 1] satisfies Propositions 10β12 for the bounded closed set Ξ© 0 0 0 0 0 Ξ©π (1) β Ξ© ; then for almost all starting points π€ = (π₯ , π ) β Μ Γ π
++ , the zero-points set π»β1 (0) = {(π₯, π, π‘) | π»(π₯, π, π‘) = Ξ© 0} defines a smooth path Ξπ€0 , which starts at (π₯0 , π0 , 1) and approaches the hyperplane π‘ = 0. Furthermore, let (π₯β , πβ , 0) be any limit point of Ξπ€0 , and π¦β = πβ π(π₯β , 0) (π(π₯β , 0) is a limit of π(π₯, π‘) as π₯ β π₯β , π‘ β 0), then π₯β is a KKT point of (1) and π¦β is the corresponding Lagrangian multiplier. Proof. Consider π» as a map of the variable (π€0 , π₯, π, π‘), for Μ Γ π
++ Γ (0, 1], which means π₯0 β Ξ©π (1)0 ; any (π€0 , π‘) β Ξ© 0 Μ , 1) < 0, because hence π(π₯ ππ» (π€0 , π₯, π, π‘) ππ€0
βπ‘πΌ 0 π =( 0 ) 0 βπ‘π (βπ₯ πΜ (π₯ , 1)) βπ‘πΜ (π₯0 , 1) (40)
is nonsingular and the Jacobi matrix of π»(π€0 , π₯, π, π‘) is of full row rank. Using the parameterized Sard theorem, Lemma 4, Μ Γ π
++ , 0 is a regular value of π»(π₯, π, π‘). for almost all π€0 β Ξ© From π₯ β π₯0 ), π» (π₯, π, 1) = ( ππΜ (π₯, 1) β π0 πΜ (π₯0 , 1)
(41)
we know that (π₯0 , π0 ) is the unique and simple solution of π»(π₯, π, 1) = 0. Then, by the implicit function theorem, there exists a smooth curve Ξπ€0 β Ξ©π (π‘)0 Γπ
++ Γ(0, 1] starting from (π₯0 , π0 , 1) and being transversal to the hyperplane π‘ = 1. Since Ξπ€0 can be extended in Ξ©π (π‘)0 Γ π
++ Γ (0, 1) until it converges to the boundary of Ξ©π (π‘)0 Γ π
++ Γ (0, 1), there must exist an extreme point (π₯, π, π‘) β π(Ξ©π (π‘) Γ π
+ Γ [0, 1]). Let (π₯, π, π‘) be any extreme point of Ξπ€0 other than (π₯0 , π0 , 1); then only the following five cases are possible: (1) (π₯, π, π‘) β Ξ© Γ π
+ Γ {0}, π < +β;
Abstract and Applied Analysis
7 Let (π₯β , πβ , π(π₯β , 0)) be any accumulation point; π¦πβ = πβ π π (π₯β , 0). By (39) and the fact that π(π₯, π‘) β 0 as π‘ β 0, we have
(2) (π₯, π, π‘) β Ξ©π (1) Γ π
+ Γ {1}, π < +β; (3) (π₯, π, π‘) β πΞ©π (π‘) Γ π
++ Γ (0, 1), π < +β; (4) (π₯, π, π‘) β Ξ©π (π‘) Γ {0} Γ (0, 1);
π
βπ (π₯β ) + βπ¦πβ βππ (π₯β ) = 0.
(5) (π₯, π, π‘) β Ξ©π (π‘) Γ {+β} Γ [0, 1]. Case (2) means that π»(π₯, π, 1) = 0 has another solution (π₯, π) except (π₯0 , π0 ), or (π₯0 , π0 ) is a double solution of π»(π₯, π, 1) = 0, which contradicts the fact that (π₯0 , π0 ) is the unique and simple solution of π»(π₯, π, 1) = 0. Case (3) means Μ π‘) = 0, 0 < π < β, and 0 < π‘ < 1; hence π(π₯, ππΜ (π₯, π‘) β π‘π0 πΜ (π₯0 , 1) = βπ‘π0 πΜ (π₯0 , 1) < 0,
(42)
which contradicts the last equation of (39). Case (4) means Μ π‘) β€ 0, π = 0, and 0 < π‘ < 1; hence π(π₯, ππΜ (π₯, π‘) β π‘π0 πΜ (π₯0 , 1) = βπ‘π0 πΜ (π₯0 , 1) < 0,
(43)
which contradicts the last equation of (39). Then, Cases (2), (3), and (4) are impossible. Because π»(π₯, π, 1) = 0 has only one and also unique solution (π₯0 , π0 ), Case (2) is impossible. By the continuity and the last equation of (39), we know that Cases (3) and (4) are impossible. If Case (5) holds, there must exist a sequence π β π₯, ππ β β, {(π₯π , ππ , π‘π )}β π=1 on Ξπ€0 such that π₯ and π‘π β π‘. By the last equation of (39), we have πΜ (π₯π , π‘π ) =
π‘π π0 πΜ (π₯0 , 1) ππ
σ³¨β 0
(44)
as π β β, which means π₯ β πΞ©π (π‘). The following two cases may be possible. (1) If π‘ = 1, let πΌ be any accumulation point of {(1 β π‘π )ππ }, which is known to exist from (39) and Μ π , π‘π ) =ΜΈ 0; then we have limπββ βπ₯ π(π₯ π₯0 = π₯ + πΌβπ₯ πΜ (π₯, 1) .
(45)
If πΌ = 0, then π₯ = π₯0 contradicts π₯0 β Ξ©π (1)0 ; else contradicts Proposition 12.
(47)
π=1
By the fact that limπ‘β0 Ξ©π (π‘) = Ξ©, we have π₯β β Ξ©. If π₯β β Ξ© , we know that 0
lim πΜ (π₯, π‘)
π₯σ³¨βπ₯β ,π‘β0
β€
lim
π₯σ³¨βπ₯β ,π‘β0
(max {πmax (π₯) , βπ (π‘)} + ππ‘ ln (ππΌπ + 1))
= max {πmax (π₯β ) , π (0)} = max {πmax (π₯β ) , βπΌπ2 } < 0, (48) where the first inequality comes from Proposition 9(1), the second equality comes from the continuity, and the third equality comes from the definition of π(π‘) in (11); hence πβ = 0 Μ π‘) = by the last equation of (39); else π₯β β πΞ©, limπ₯βπ₯β ,π‘β0 π(π₯, πmax (π₯β ) = 0 by Proposition 9(2) and π π (π₯β , 0) = 0 for π β πΌ(π₯β ) by Proposition 8. Thus we have π¦πβ ππ (π₯β ) = 0,
π = 1, . . . , π.
(49)
Summing up, (π₯β , π¦β ) is a solution of the KKT system of (1), which means that π₯β is a KKT point of (1) and π¦β is the corresponding Lagrangian multiplier. 3.2. The Modified Flattened Aggregate Constraint Homotopy. From Proposition 8(3), we know that π(π₯, π‘) β 0 as π‘ β 0 for any π₯ β Ξ© and, hence, we can use the following modified flattened aggregate constraint homotopy (MFACH) instead of the FACH: (1 β π‘) (βπ (π₯) + πβ (π₯, π‘)) + π‘ (π₯ β π₯0 ) ) π» (π₯, π, π‘) = ( ππΜ (π₯, π‘) β π‘π0 πΜ (π₯0 , 1) = 0,
(2) If π‘ < 1, by (39), we have
(50)
(1 β π‘) lim ππ βπ₯ πΜ (π₯π , π‘π ) = β (1 β π‘) βπ (π₯) β π‘ (π₯ β π₯0 ) ,
where
πσ³¨ββ
(46) the right-hand is finite; however, by π Μ , π‘π ) =ΜΈ 0, we know that the left-hand is limπββ βπ₯ π(π₯ infinite; this is a contradiction. As a conclusion, Case (1) is the only possible case. This implies that Ξπ€0 must approach the hyperplane π‘ = 0. π β π β Because {π(π₯π , π‘π )}β π=1 (defined in (17)), {π₯ }π=1 , and {π }π=1 are bounded sequences, we know that {(π₯π , ππ , π(π₯π , π‘π ))}β π=1 has at least one accumulation point as π β β.
β (π₯, π‘) = β π π (π₯, π‘) βππ (π₯) . πβπΌπ (π₯,π‘)
(51)
Remarks. (i) Since π(π₯, π‘) is dropped in (51), the expressions of the homotopy map (50) and its Jacobian and hence the corresponding code become simpler. Moreover, the computation of H in (50) and its Jacobi matrix DH is a little cheaper than that for the FACH method. (ii) To guarantee that π»(π₯, π, π‘) in (39) be a πΆ2 map, π(π§, π‘) must be πΆ3 ; hence, if π(π§, π‘) is chosen as a polynomial, the
8
Abstract and Applied Analysis
total degree should be seven; in contrast, for the homotopy map in (50), π(π§, π‘) is only needed to be πΆ2 . Then π(π§, π‘) should satisfy π (βπΌπ (π‘) , π‘) = 0,
ππ (βπΌπ (π‘) , π‘) = 0, ππ§
π2 π (βπΌπ (π‘) , π‘) = 0, ππ§2 π (βπ (π‘) , π‘) = 1,
ππ (βπ (π‘) , π‘) = 0, ππ§
(52)
βπ (π₯) + πβ (π₯, π‘π ) ) = 0, πΉπ‘π (π₯, π) = ( ππΜ (π₯, π‘π )
π2 π (βπ (π‘) , π‘) =0 ππ§2
6(π§ + π (π‘))5 15(π§ + π (π‘))4 + (πΌπ (π‘) β π (π‘))5 (πΌπ (π‘) β π (π‘))4 10(π§ + π (π‘))3 + + 1. (πΌπ (π‘) β π (π‘))3
(54)
where π‘π is a small positive constant. For other situations, the steplength will be decreased to make new predictor-corrector steps (see Algorithm 1).
and π(π§, π‘) can be chosen as π (π§, π‘) =
the predictor direction. The corrector step terminates when π(π,π) and π»(π€(π,π) , π‘(π,π) ) satisfy the tolerance criterions. At each predictor step and corrector step, the feasibility of the predictor point (π€π , π‘π ) and the corrector point (π€(π,π) , π‘(π,π) ) needs to be checked. If π‘π < π‘π in the predictor step or π‘π < 0 in the corrector step, a damping step is used to get a new point (π€(π,0) , 0). Then, if π€(π,0) is feasible, the end game, a more efficient strategy than predictor-corrector steps when the homotopy parameter is close to 0, is invoked. Starting with π€(π,0) , Newtonβs method is used to solve
(53)
(iii) The existence and convergence of the homotopy path defined by the MFACH can be proven in a similar way with that of the FACH.
4. The FACH-S-N Procedure and Numerical Results 4.1. The FACH-S-N Procedure. In this section, we give a numerical procedure, FACH-S-N procedure, to trace the flattened aggregate constraint homotopy path by secant predictor and Newton corrector steps. It consists of three main steps: the predictor step, the corrector step, and the end game. The predictor step is an approximate step along the homotopy path: it uses a predictor direction ππ and a steplength βπ to get a predictor point. The first predictor direction uses the tangent direction, and others use the secant direction. The steplength is determined by several parameters. It is set to no more than 1, which can ensure that the predictor point is close to the homotopy path and hence stays in the convergence domain of Newtonβs method in the corrector step. If the angle of the predictor direction and the previous one π½π is greater than π/4, the steplength will be decreased to avoid using the opposite direction as the predictor direction. If the corrector criteria are satisfied with no more than four Newton iterations for three times in succession, the steplength will be increased or kept invariable. Otherwise, the steplength will be decreased. Once a predictor point (π€(π,0) , π‘(π,0) ) is calculated, one or more Newton iterations are used to bring the predictor point back to the homotopy path in the corrector step. The corrector points (π€(π,π) , π‘(π,π) ), π = 1, . . . , 5, are calculated by (π€(π,π+1) , π‘(π,π+1) ) = (π€(π,π) , π‘(π,π) ) + π(π,π+1) , π = 0, . . . , 4, where the step π(π,π+1) is the solution of an augmented system defined by the homotopy equation and the direction perpendicular to
4.2. The Numerical Experiment. Although there exist so many test problems, such as the CUTEr test set [21], we cannot find a large collection of test problems with moderate variables and very many complicated nonlinear constraints. In this paper, six test problems are chosen to test the algorithm. Problem 4.1 is chosen from the CUTEr test set and it is used to illustrate a special situation; others are derived from the discretized semi-infinite programming problems. We also give two artificial test problems 4.2 and 4.3 and use three problems 4.4β4.6 in [22]. The numbers of variables in problem 4.2 and the number of constrains in problems 4.2β 4.6 can be arbitrary. For each test problem, the gradients and Hessians of the objective and constraint functions are evaluated analytically. The FACH-S-N procedure was implemented in MATLAB. To illustrate its efficiency, we also implemented the ACH method in [19] using a similar procedure. In addition, we downloaded the state-of-the-art solver KNITRO, which provides three algorithms for solving large-scale nonlinear programming problems, and we used the interior-point direct method with default parameters to compare with the FACH and MFACH methods. For any iterate point π₯π , if πmax (π₯π ) < 10β6 , it was treated as a feasible point. The test results were obtained by running MATLAB R2008a on a desktop with Windows XP Professional operation system, Intel(R) Core(TM) i5-750 2.66 GHz processor, and 8 GB of memory. The default parameters were chosen as follows. (i) Parameters for the flattened aggregate constraint function: π = 0.01, π1 = 0.05, π2 = 0.5 β 10β5 , and πΌ = 2. (ii) Parameters in the end game section: π‘π = 10β6 and π‘end = 0.1. (iii) Step size parameters: β0 = 0.1, π΅min = [0.5, 0.75], and π΅max = [3, 1.5]. (iv) Tracking tolerances: π»tol = 10β5 and π»final = 10β12 . (v) Initial Lagrangian multipliers: 1 for ACH, FACH, and MFACH methods.
Abstract and Applied Analysis
9
Μ Γ π
++ ; initial steplength β1 , steplength Input Give π β (0, 1], π1 > 0, π2 > 0, πΌ > 1, π‘end and π‘π ; starting point π€0 β Ξ© contraction factor π΅min , steplength expansion factor π΅max ; tracking tolerances π»tol and π»final for correction. Set π‘0 = 1, π = 1, IT = 0, πgood = 2, π½1 = 0. Step 1. The predictor step. Step 1.1. Compute the predictor direction. If π = 1, compute π by solving π»σΈ (π₯0 , π0 , 1) ) π = βπ0 , ( π (π0 ) where π0 = (0, . . . , 0, β1) β π
π+2 , set the predictor direction π1 = π/ βπβ; else, compute the predictor direction ππ , the angle π½π between ππ and ππβ1 : π πβ1 (π€πβ1 , π‘πβ1 ) β (π€πβ2 , π‘πβ2 ) ππ = , π½π = arccos ((ππ ) π ) . πβ1 πβ2 β(π€ , π‘πβ1 ) β (π€ , π‘πβ2 )β Step 1.2. Adjust the steplength. Adjust the steplength βπ as follows: if the predictor point is infeasible, set βπ = π΅min (1)βπ , πgood = 0; else if the corrector step fails or π½π > π/4, set βπ = π΅min (1)βπβ1 , πgood = 0; else if π = 5, set βπ = π΅min (2)βπβ1 , πgood = 0; else if π = 4, set βπ = βπβ1 , πgood = πgood + 1; else if π = 3, set πgood = πgood + 1, if πgood > 2, set βπ = min{1, π΅max (2)βπβ1 }, else, set βπ = βπβ1 ; else, set πgood = πgood + 1, if πgood > 2, set βπ = min{1, π΅max (1)βπβ1 }, else, set βπ = βπβ1 . If βπ < 10β10 , stop the algorithm with an error flag. Step 1.3. Compute the predictor point and check its feasibility. Compute the predictor point (π€(π,0) , π‘(π,0) ) = (π€πβ1 , π‘πβ1 ) + βπ ππ . 0 (π,0) β Ξ©π (π‘(π,0) ) Γ π
++ or π‘(π,0) β₯ 1, If π€ the predictor point is infeasible, goto Step 1.2; else if π‘(π,0) β€ π‘end , adjust the steplength βπ , compute the point (π€(π,0) , 0), if π€(π,0) β Ξ© Γ π
+ , goto Step 3, else, the predictor point is infeasible, goto Step 1.2; else, goto Step 2. Step 2. The corrector step. Set π = 0. Repeat If π = 5, set IT = IT + 5, π€π = π€πβ1 , π‘π = π‘πβ1 , ππ+1 = ππ , π = π + 1, the corrector step fails, goto Step 1.2. Compute the Newton step π(π,π+1) by solving π»σΈ (π€(π,π) , π‘(π,π) ) (π,π+1) βπ» (π€(π,π) , π‘(π,π) ) )π =( ), ( π π 0 (π ) the corrector point (π€(π,π+1) , π‘(π,π+1) ) = (π€(π,π) , π‘(π,π) ) + π(π,π+1) ; set π = π + 1. If π€(π,π) β Ξ©π (π‘(π,π) )0 Γ π
++ or π‘(π,π) β₯ 1, set IT = IT + π, π€π = π€πβ1 , π‘π = π‘πβ1 , ππ+1 = ππ , π = π + 1, the corrector step fails, goto Step 1.2; else if π‘(π,π) < 0, compute the point (π€(π,0) , 0) by a damping Newton step, set IT = IT + π, if π€(π,0) β Ξ© Γ π
+ , goto Step 3, else, set π€π = π€πβ1 , π‘π = π‘πβ1 , ππ+1 = ππ , π = π + 1, the corrector step fails, goto Step 1.2. (π,π) Until βπ»(π€ , π‘(π,π) )βinf β€ π»tol and βπ(π,π) β β€ π»tol . If π‘π < π‘π , return with π₯β = π₯π , stop the algorithm; else, set π»tol = min{π»tol , π‘π }, π = π + 1, goto Step 1. Step 3. The end game. Set π = 0. Repeat If π = 5 or βπ(π,π) β > βπ(π,πβ1) β for π > 1, or π€(π,π) β Ξ© Γ π
+ , set π‘end = 0.3 β π‘end , π€π = π€πβ1 , π‘π = π‘πβ1 , ππ+1 = ππ , IT = IT + π, π = π + 1, the corrector step fails, goto Step 1.2. Compute the Newton step π(π,π+1) by solving πΉπ‘σΈ π (π€(π,π) )π(π,π+1) = βπΉπ‘π (π€(π,π) ), the corrector point π€(π,π+1) = π€(π,π) + π(π,π+1) ; Algorithm 1: Continued.
10
Abstract and Applied Analysis
set π = π + 1. Until βπΉπ‘π (π€(π,π) )βinf β€ π»final and βπ(π,π) β β€ π»final . Set IT = IT + π, return with π₯β = π₯(π,π) , stop the algorithm. Algorithm 1: The FACH-S-N procedure.
For each problem with different parameters, we list the value of objective function π(π₯) and the maximal function of constraints πmax (π₯) at π₯β , the number of iterations πΌπ, the number of evaluations of gradients of individual constraint functions in the FACH and MFACH methods ππ (in contrast, ππ is π β πΌπ in the ACH method), and the running time in seconds ππππ. For problems that were not solved by the conservative setting, we also give the reason for failure. The notation βfail1 β indicates that the steplength in predictor step is smaller than 10β10 ; it is generally due to poor conditioned Jacobian matrix. The notation βfail2 β means out of memory. The notation βfail3 β means no result in 5000 iterations or 3600 seconds.
Table 1: Test results for Example 14. Method ACH KNITRO FACH MFACH
2 ), β 100 sin (βπ₯π + 1.5π + π₯πβ1
(55)
+
(π‘πσΈ β π₯2 ) π₯42
β 1,
π‘π =
π , (βπ β 1)
π = 0, . . . , βπ β 1,
π‘πσΈ =
π , (βπ β 1)
π = 0, . . . , βπ β 1,
2β€πβ€1000
(57)
π₯0 = (0, 0, 100, 100) β π
4 . Example 17. Consider
π₯0 = (1, . . . , 1) β π
1000 . Remark 15. In this example, the starting point happens to be an unconstrained minimizer of the objective function and we found that the π₯-components of iterative points generated by the FACH and MFACH methods (not other homotopy methods) remain invariant. This is not an occasional phenomenon. In fact, if (π₯0 , 0) β Ξ©0 Γ π
+π is a solution of the KKT system, that is, π₯0 is a stationary point of the objective function π(π₯), when the parameters πΌ, π1 , and π2 in FACH and MFACH methods satisfy πmax (π₯0 ) β€ βπΌ(π1 + π2 ), the π₯-components of points on the homotopy path Ξπ€0 will remain invariant, which can be derived from the fact that (π₯0 , π‘π0 π(1)/π(π‘), π‘) is the solution of (39) and (50) for π‘ β (0, 1]. Moreover, since π»σΈ (π₯0 , π0 , 1) = (
Time 85.5496 2.8733 8.3912 8.4150
2
2
(π‘ β π₯ ) ππ,π (π₯) = π 2 1 π₯3
π = 1; π = 2, . . . , 1000; π = 1001; π = 1002, . . . , 2000,
ππ β β 0 0
IT 35 0 4 4
π (π₯) = π₯32 + π₯42 ,
π (π₯) = sin (π₯1 β 1 + 1.5π)
π₯1 β π { { 2 { {π₯πβ1 β π₯π β π ππ (π₯) = { { βπ₯ β π { { 12 {βπ₯πβ1 + π₯π β π
πmax (π₯β ) β2.1416 β2.1416 β2.1416 β2.1416
Example 16. Consider
Example 14 (see [21]). Consider
+
π(π₯β ) β99901.0000 β99901.0000 β99901.0000 β99901.0000
πΌ 0 0 ), 0 πΜ (π₯0 , 1) β
(56)
the π₯-component of π1 is 0. By a similar discussion, the π₯component of π(1,π) is 0, and hence π₯1 = π₯(1,π) = π₯0 . In this analogy, we know that π₯π = π₯(π,π) = π₯0 for any π and π in algorithm FACH-S-N.
π (π₯) =
1 2 β (π₯ β 1) , π 1β€πβ€π π
ππ (π₯) = β cos (π‘π π₯π ) + π‘π β π₯π3 , π‘π =
1β€πβ€π
1β€πβ€π
0.5 + ππ , (π β 1)
π = 0, . . . , π β 1,
(58)
π₯0 = (β2, . . . , β2) β π
π . Example 18 (see [22]). Consider π (π₯) = π₯12 + π₯22 + π₯32 , ππ (π₯) = π₯1 + π₯2 exp (π₯3 π‘π ) + exp (2π‘π ) β 2 sin (4π‘π ) , π‘π =
π , (π β 1)
π = 0, . . . , π β 1,
π₯0 = (β200, β200, 200) β π
3 .
(59)
Abstract and Applied Analysis
11 Table 2: Test results for Example 16.
π
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
102
104
106
β
π(π₯ ) 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
πmax (π₯β ) β1.3863e β 08 β7.1233e β 10 β1.3863e β 08 β1.3863e β 08 β1.3863e β 08 β8.0005e β 10 β1.3863e β 08 β1.3863e β 08 β1.3863e β 08 β7.8830e β 10 β1.3863e β 08 β1.3863e β 08
IT 394 18 541 567 397 20 541 567 404 13 541 567
ππ β β 721 823 β β 721 823 β β 805 907
Time 0.2423 0.2288 0.3270 0.2692 1.9730 0.7227 0.4359 0.4676 249.3395 2170.4301 10.9486 11.2815
Table 3: (a) Test results for Example 17 with π = 100. (b) Test results for Example 17 with π = 500. (c) Test results for Example 17 with π = 1000. (d) Test results for Example 17 with π = 2000. (a)
π 102
103
104
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
β
π(π₯ ) 1.4918 0.2603 1.4915 1.4915 1.4918 0.2603 1.4915 1.4915 1.4918 0.2603 1.4915 1.4915
πmax (π₯β ) β1.5282e β 03 β1.9708e β 06 β6.6613e β 16 6.1062e β 15 β1.5282e β 03 β2.7042e β 05 β6.6613e β 16 6.1062e β 15 β1.5282e β 03 β2.5434e β 04 β6.6614e β 16 6.1062e β 15
IT 243 32 156 144 243 19 156 144 243 18 156 144
ππ β β 96 85 β β 96 85 β β 96 86
Time 36.9898 4.1262 0.4067 0.3186 325.6083 25.1720 0.6852 0.5342 3218.5422 271.1651 5.1941 4.0120
IT 405 31 159 139 β β 159 139 β β 159 139
ππ β β 45 25 β β 45 25 β β 45 25
Time 1371.0198 95.1782 3.9951 2.9676 fail3 fail3 6.5173 4.6898 fail2 fail2 31.7107 25.1842
IT β 27 197 197
ππ β β 46 46
Time fail3 359.4507 20.8919 21.2641
(b)
π 102
103
104
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
β
π(π₯ ) 1.2524 0.1417 1.2510 1.2510 β β 1.2510 1.2510 β β 1.2510 1.2510
πmax (π₯β ) β1.0693e β 02 β2.0927e β 05 1.2490e β 14 1.2490e β 14 β β 1.2490e β 14 1.2490e β 14 β β 1.2490e β 14 1.2490e β 14 (c)
π 102
Method ACH KNITRO FACH MFACH
π(π₯β ) β 0.1099 1.1880 1.1880
πmax (π₯β ) β β5.6905e β 05 2.3370e β 14 2.4258e β 14
12
Abstract and Applied Analysis (c) Continued.
π
β
103
104
πmax (π₯β ) β β 2.3370e β 14 2.4258e β 14 β β 2.3370e β 14 2.4258e β 14
π(π₯ ) β β 1.1880 1.1880 β β 1.1880 1.1880
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
IT β β 197 197 β β 197 197
ππ β β 46 46 β β 46 46
Time fail2 fail2 28.8674 27.5140 fail2 fail2 101.3199 99.1354
IT β β 260 258 β β 260 258 β β 260 258
ππ β β 47 46 β β 47 46 β β 47 46
Time fail3 fail3 126.6290 125.7973 fail2 fail2 146.6079 145.7293 fail2 fail2 368.5331 369.3701
IT 1605 2 815 820 1605 2 815 820 1606 2 960 931
ππ β β 288 293 β β 288 294 β β 6444 7094
(d)
π 102
103
104
β
πmax (π₯β ) β β β6.2839e β 14 4.0967e β 14 β β β6.2839e β 14 4.0967e β 14 β β β6.2839e β 14 4.0967e β 14
π(π₯ ) β β 1.1406 1.1406 β β 1.1406 1.1406 β β 1.1406 1.1406
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
Table 4: Test results for Example 18. π
π(π₯β ) 5.3347 1.2000e + 05 5.3347 5.3347 5.3347 1.2000e + 05 5.3347 5.3347 5.3347 1.2000e + 05 5.3347 5.3347
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
102
104
106
πmax (π₯β ) β4.4409e β 16 β3.9900e + 02 1.3323e β 15 1.3323e β 15 β4.4409e β 16 β3.9900e + 02 1.3323e β 15 1.3323e β 15 1.3323e β 15 β3.9900e + 02 β9.2480e β 08 β1.1278e β 07
Example 19 (see [22]). Consider
Example 20 (see [22]). Consider
π (π₯) = (π₯1 β 2π₯2 + 5π₯22 β π₯23 β 13) + (π₯1 β 14π₯2 +
π₯22
+
π₯23
π (π₯) =
2
π , (π β 1)
π₯0 = (0, β45) β π
2 .
2
β 29) ,
π = 0, . . . , π β 1,
π₯12 π₯1 + + π₯22 , 3 2
ππ (π₯) = (1 β π₯12 π‘π2 ) β π₯1 π‘π2 β π₯22 + π₯2 ,
2
ππ (π₯) = π₯12 + 2π₯2 π‘π2 + exp (π₯1 + π₯2 ) β exp (π‘π ) , π‘π =
Time 0.7546 0.0141 0.2977 0.3004 7.0581 0.3163 1.3127 1.3185 575.4663 3243.2201 42.7664 41.7933
(60)
π π‘π = , (π β 1)
(61)
π = 0, . . . , π β 1,
π₯0 = (β1, 100) β π
2 . To explain the numerical efficiency, we make the following remarks by numerical results in Tables 1, 2, 3, 4, 5 and 6.
Abstract and Applied Analysis
13 Table 5: Test results for Example 19.
π 102
104
106
β
Method ACH KNITRO FACH MFACH ACH KNITRO FACH MFACH
π(π₯ ) 97.1589 97.1592 97.1589 97.1589 97.1589 102.2095 97.1589 97.1589
πmax (π₯β ) 0.0000 β7.5266 β 05 0.0000 0.0000 0.0000 β6.7465 β 01 0.0000 0.0000
IT 213 68 170 160 227 263 249 251
ππ β β 191 183 β β 10926 11321
Time 0.0964 0.5162 0.0795 0.0608 0.5310 15.4868 0.2846 0.2783
ACH KNITRO FACH MFACH
97.1589 188.9496 97.1589 97.1589
0.0000 β8.3841e β 01 0.0000 0.0000
229 17 334 253
β β 427520 486506
50.8763 2206.4660 9.5724 6.8910
Table 6: Test results for Example 20. Method
π(π₯β )
πmax (π₯β )
IT
ππ
Time
10
ACH KNITRO FACH MFACH
2.4305 2.4305 2.4305 2.4305
0.0000 β1.3822e β 10 2.2204e β 16 2.2204e β 16
342 21 348 341
β β 387 374
0.1749 0.0291 0.1368 0.1276
104
ACH KNITRO FACH MFACH
2.4305 2.4305 2.4305 2.4305
β6.6632e β 09 β1.6180e β 10 β6.6632e β 09 β6.6632e β 09
344 38 296 358
β β 14710 23559
0.9804 1.0329 0.2610 0.2802
106
ACH KNITRO FACH MFACH
2.4305 2.6182 β 2.4305
β4.9783e β 08 β1.2609e β 04 β β4.9021e β 08
344 61 β 516
β β β 2231308
94.3943 2372.2093 fail1 16.3756
π 2
(i) If πmax (π₯) β€ βπΌπ(π‘) for any (π₯, π, π‘) β Ξπ€0 , the FACH and MFACH methods do not need to calculate the gradient and Hessian of any constraint functions. (ii) For problems whose gradients and Hessians of constraint functions are expensive to evaluate, the performance of the FACH and MFACH methods is much better than the ACH method based on similar numerical tracing procedures. (iii) Compared to the interior-point direct algorithm of the state-of-the-art solver KNITRO, the test results show that the FACH and MFACH methods perform worse when π is small but much better when π is large for most problems. In addition, we can see that their time cost increases more slowly than KNITRO solver as π increases. (iv) The function π(π§, π‘) is important for the flattened aggregate constraint function. Theoretically, the parameters π1 , π2 , and πΌ in the FACH and MFACH methods can be chosen freely. However, they do matter in the practical efficiency of the FACH and MFACH methods and should be suitably chosen. If
these parameters are too large, then too many gradients and Hessians of individual constraint functions need to be evaluated and, hence, cause low efficiency. On the other hand, if they are too small, the Hessian of the flattened aggregated constraint function (10) may become ill-conditioned. In our numerical tests, we fixed π1 = 0.05, π2 = 0.5 β 10β5 , and πΌ = 2. In addition, the function π(π§, π‘) can be defined in many ways, and preliminary numerical experiments show that the algorithms with different functions π(π§, π‘) have similar efficiencies. (v) In algorithm FACH-S-N, we gave only a simple implementation of the ACH, FACH, and MFACH methods. To improve implementation of the FACH and MFACH methods, a lot of work needs to be done on all processes of numerical path tracing, say, schemes of predictor and corrector, steplength updating, linear system solving, and end game. Other practical strategies in the literature for large-scale nonlinear programming problems (e.g., [9, 23β26]) are also very important for improving the efficiency.
14
Abstract and Applied Analysis
5. Conclusions By introducing a flattened aggregate constraint function, a flattened aggregate constraint homotopy method is proposed for nonlinear programming problems with few variables and many nonlinear constraints, and its global convergence is proven. By greatly reducing the computation of gradients and Hessians of constraint functions in each iteration, the proposed method is very competitive for nonlinear programming problems with a large number of complicated constraint functions.
Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper.
[12]
[13]
[14]
[15]
[16]
Acknowledgment The research was supported by the National Natural Science Foundation of China (11171051, 91230103).
[17]
References
[18]
[1] A. Ben-Tal and A. Nemirovski, Lectures on Modern Convex Optimization: Analysis, Algorithms, and Engineering Applications, Society for Industrial and Applied Mathematics, Philadelphia, Pa, USA, 2001. [2] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, Cambridge, UK, 2004. [3] C. B. Garcia and W. I. Zangwill, Pathways to Solutions, Fixed Points, and Equilibria, Prentice Hall, Englewood Cliffs, NJ, USA, 1981. [4] Y. Nesterov and A. Nemirovskii, Interior-Point Polynomial Algorithms in Convex Programming, vol. 13 of SIAM Studies in Applied Mathematics, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, Pa, USA, 1994. [5] T. Terlaky, Interior Point Methods of Mathematical Programming, vol. 5 of Applied Optimization, Kluwer Academic Publishers, Dordrecht, The Netherlands, 1996. [6] Y. Y. Ye, Interior Point Algorithms, Wiley-Interscience Series in Discrete Mathematics and Optimization, John Wiley & Sons, New York, NY, USA, 1997. [7] A. Forsgren and P. E. Gill, βPrimal-dual interior methods for nonconvex nonlinear programming,β SIAM Journal on Optimization, vol. 8, no. 4, pp. 1132β1152, 1998. [8] D. M. Gay, M. L. Overton, and M. H. Wright, βA primal-dual interior method for nonconvex nonlinear programming,β in Advances in Nonlinear Programming, vol. 14 of Applied Optimization, pp. 31β56, Kluwer Academic Publishers, Dordrecht, The Netherlands, 1998. [9] R. J. Vanderbei and D. F. Shanno, βAn interior-point algorithm for nonconvex nonlinear programming,β Computational Optimization and Applications, vol. 13, no. 1β3, pp. 231β252, 1999. [10] G. Feng, Z. Lin, and B. Yu, βExistence of an interior pathway to a Karush-Kuhn-Tucker point of a nonconvex programming problem,β Nonlinear Analysis: Theory, Methods & Applications, vol. 32, no. 6, pp. 761β768, 1998. [11] G. C. Feng and B. Yu, βCombined homotopy interior point method for nonlinear programming problems,β in Advances in
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
Numerical Mathematics, vol. 14 of Lecture Notes in Numerical and Applied Analysis, pp. 9β16, 1995. L. T. Watson, βTheory of globally convergent probability-one homotopies for nonlinear programming,β SIAM Journal on Optimization, vol. 11, no. 3, pp. 761β780, 2000/01. Q. H. Liu, B. Yu, and G. C. Feng, βAn interior point pathfollowing method for nonconvex programming with quasi normal cone condition,β Advances in Mathematics, vol. 19, no. 4, pp. 281β282, 2000. B. Yu, Q. H. Liu, and G. C. Feng, βA combined homotopy interior point method for nonconvex programming with pseudo cone condition,β Northeastern Mathematical Journal, vol. 16, no. 4, pp. 383β386, 2000. Y. F. Shang, Constraint shifting combined homotopy method for nonlinear programming, equilibrium programming and variational inequalities [Ph.D. thesis], Jilin University, 2006. B. Yu and Y. F. Shang, βBoundary moving combined homotopy method for nonconvex nonlinear programming,β Journal of Mathematical Research and Exposition, vol. 26, no. 4, pp. 831β 834, 2006. X. S. Li, βAn aggregate function method for nonlinear programming,β Science in China A: Mathematics, Physics, Astronomy, vol. 34, no. 12, pp. 1467β1473, 1991. B. W. Kort and D. P. Bertsekas, βA new penalty function algorithm for constrained minimization,β in Proceedings of the 1972 IEEE Conference on Decision and Control, New Orleans, La, USA, 1972. B. Yu, G. C. Feng, and S. L. Zhang, βThe aggregate constraint homotopy method for nonconvex nonlinear programming,β Nonlinear Analysis: Theory, Methods & Applications, vol. 45, no. 7, pp. 839β847, 2001. E. L. Allgower and K. Georg, Introduction to Numerical Continuation Methods, vol. 45 of Classics in Applied Mathematics, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, Pa, USA, 2003. N. I. M. Gould, D. Orban, and P. L. Toint, βCUTEr (and sifdec), a constrained and unconstrained testing environment, revisited,β Tech. Rep. TR/PA/01/04, CERFACS, Toulouse, France, 2001. D. H. Li, L. Qi, J. Tam, and S. Y. Wu, βA smoothing Newton method for semi-infinite programming,β Journal of Global Optimization, vol. 30, no. 2-3, pp. 169β194, 2004. R. H. Byrd, M. E. Hribar, and J. Nocedal, βAn interior point algorithm for large-scale nonlinear programming,β SIAM Journal on Optimization, vol. 9, no. 4, pp. 877β900, 1999. A. R. Conn, N. I. M. Gould, and Ph. L. Toint, LANCELOT: A Fortran Package for Large-Scale Nonlinear Optimization (Release A), vol. 17 of Springer Series in Computational Mathematics, Springer, New York, NY, USA, 1992. R. Fletcher and S. Leyffer, βNonlinear programming without a penalty function,β Mathematical Programming A, vol. 91, no. 2, pp. 239β269, 2002. P. E. Gill, W. Murray, and M. A. Saunders, βSNOPT: an SQP algorithm for large-scale constrained optimization,β SIAM Journal on Optimization, vol. 12, no. 4, pp. 979β1006, 2002.
Advances in
Operations Research Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Advances in
Decision Sciences Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Applied Mathematics
Algebra
Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Probability and Statistics Volume 2014
The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of
Differential Equations Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014
Submit your manuscripts at http://www.hindawi.com International Journal of
Advances in
Combinatorics Hindawi Publishing Corporation http://www.hindawi.com
Mathematical Physics Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Complex Analysis Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of Mathematics and Mathematical Sciences
Mathematical Problems in Engineering
Journal of
Mathematics Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Discrete Mathematics
Journal of
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Discrete Dynamics in Nature and Society
Journal of
Function Spaces Hindawi Publishing Corporation http://www.hindawi.com
Abstract and Applied Analysis
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of
Journal of
Stochastic Analysis
Optimization
Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014